Seerie: An Offline 1.7B-Parameter Assistant for Scam Prevention in India, and a Manual Audit of Its Failures
Jay Tiwari, Seerror Technologies, Bengaluru. Fine-tuned Qwen3-1.7B with QLoRA on 36,669 records; deployed as a 1.11 GB GGUF model. Manual audit of 70 prompts: 58 safe (82.9%), 12 wrong. Automatic harness 90%, Cohen's κ = 0.10. Errors concentrate in fabricated legal sections and Hinglish prompts.
Views —
Downloads —
Read paper →
PDF
Zenodo DOI